A Guardian investigation published August 26 pulled apart a US-based outfit calling itself the Hanover Institute for Public Policy. There's nothing behind it: no corporate registration in any jurisdiction, no office address on the site, no staff list, and not a single bylined author on any of its reports.
What it does have is output. Over nine days, from August 6 to 14, the site published 124 reports totaling more than 500,000 words, all taking a pro-Israel position. The topics ranged across genocide allegations, war crimes, the torture of Palestinian detainees, and starvation conditions in Gaza — precisely the questions international human rights bodies are actively investigating.
Academic Dressing, Written for Models
The reports are formatted and worded to look like policy research: neutral tone, footnotes, sourcing. The Guardian identified a recurring technique — in 20 of the reports, whenever a human rights finding on genocide or apartheid came up, the text hedged with the same line: no court has yet ruled on these allegations. The underlying facts aren't disputed; what gets diluted is how much weight the conclusion carries.
Researcher Nick Cleveland-Stout's assessment was short.
"It succeeds in masquerading as very academic."
What separates this from earlier social-media influence campaigns is who it's aimed at. The site runs on a commercial service called AI Story Optimization, whose pitch is blunt: optimize content so systems like ChatGPT, Perplexity, Claude, and Gemini are more likely to cite you. Human readers barely matter here — the product is getting scraped by a model and landing in its list of cited sources.
SEO's Move to AEO, in Practice
Search-era content farms won by stacking keywords and backlinks to chase the top ten results. This version has switched targets. When a model synthesizes an answer, it looks for sources that read as authoritative, structured, and properly cited — and a site dressed up as a think tank checks every one of those boxes. That explains the cadence: 500,000 words in nine days is about coverage and consistency, not about anyone actually reading it.
What makes this hard to catch is that none of it is fabricated news. Most of the facts in the reports check out; what's been manipulated is the framing and the weighting. Fact-checking tools are mostly useless against this kind of material, because the checkable parts are all true.
A Filing Put the Money Trail on the Record
The story got confirmed through a compliance filing. Piro Inc, a New York media production company, filed with the US Department of Justice this month under the Foreign Agents Registration Act, registering the material as content distributed on behalf of the Israeli government. That 1938 law was built to require disclosure from people working for foreign governments — originally aimed at newspaper ads and lobbying — and it's now being used to label a batch of content written for models.
Piro co-founder Rosenberg described the work as being hired to "place accurate, sourced facts." That description doesn't actually contradict what the Guardian found — the two claims describe the same operation from opposite sides.
For the companies building these models, the problem sits in retrieval. A domain that's new, has no legal entity, no named authors, and produced 500,000 words in nine days should score low on traditional authority signals. But when a model's retrieval ranking favors structure and citation formatting, those same features flip into an advantage. Do the math: 124 reports over nine days works out to a dozen-plus reports and more than 40,000 words a day — an output no real think tank could sustain, but routine for a content pipeline.
A cluster of companies already sell "AI visibility" as a legitimate service, mostly to brands that want to get recommended. What's different here is that the same toolkit got pointed at a foreign-policy narrative, and the paper trail exists this time because the law required it. Next time, it might not.
Sources: The Guardian investigation, Arab News, Cybernews, CocoLoop, US Department of Justice FARA filings; report counts, word totals, and the filing entity were cross-checked against public reporting.